Marine object detection in forward-looking sonar images via semantic-spatial feature enhancement

Forward-looking sonar object detection plays a vital role in marine applications such as underwater navigation, surveillance, and exploration, serving as an essential underwater acoustic detection method. However, the challenges posed by seabed reverberation noise, complex marine environments, and varying object scales significantly hinder accurate detection of diverse object categories. To overcome these challenges, we propose a novel semantic-spatial feature enhanced detection model, namely YOLO-SONAR, tailored for marine object detection in forwardlooking sonar imagery. Specifically, we introduce the competitive coordinate attention mechanism (CCAM) and the spatial group enhance attention mechanism (SGEAM), both integrated into the backbone network to effectively capture semantic and spatial features within sonar images, while feature fusion is employed to suppress complex marine background noise. To address the detection of small-scale marine objects, we develop a context feature extraction module (CFEM), which enhances feature representation for tiny object regions by integrating multi-scale contextual information. Furthermore, we adopt the Wise-IoUv3 loss function to mitigate the issue of class imbalance within marine sonar datasets and stabilize the model training process. Experimental evaluations conducted on real-world forward-looking sonar datasets, MDFLS and WHFLS, demonstrate that the proposed detection model outperforms other state-of-the-art methods, achieving an average precision (mAP) of 81.96% on MDFLS and 82.30% on WHFLS, which are improvements of 7.65% and 12.89%, respectively, over the best-performing existing methods. These findings highlight the potential of our approach to significantly advance marine object detection technologies, facilitating more efficient underwater exploration and monitoring.